Churn Prediction in Online Games Using Players’ Login Records: A Frequency Analysis Approach

The rise of free-to-play and other service-based business models in the online gaming market brought to game publishers problems usually associated to markets like mobile telecommunications and credit cards, especially customer churn. Predictive models have long been used to address this issue in these markets, where companies have a considerable amount of demographic, economic, and behavioral data about their customers, while online game publishers often only have behavioral data. Simple time series’ feature representation schemes like RFM can provide reasonable predictive models solely based on online game players’ login records, but maybe without fully exploring the predictive potential of these data. We propose a frequency analysis approach for feature representation from login records for churn prediction modeling. These entries (from real data) were converted into fixed-length data arrays using four different methods, and then these were used as input for training probabilistic classifiers with the $k$ -nearest neighbors machine learning algorithm. The classifiers were then evaluated and compared using predictive performance metrics. One of the methods, the time-frequency plane domain analysis, showed satisfactory results, being able to theoretically increase the retention campaigns profits in more than 20% over the RFM approach.